On 06/03/2021 18:13, Bart Van Assche wrote:
On 3/5/21 7:14 AM, John Garry wrote:
@@ -2296,10 +2296,14 @@ void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct 
blk_mq_tags *tags,
for (i = 0; i < tags->nr_tags; i++) {
                        struct request *rq = tags->static_rqs[i];
+                       int j;
if (!rq)
                                continue;
                        set->ops->exit_request(set, rq, hctx_idx);
+                       /* clean up any references which occur in @ref_tags */
+                       for (j = 0; ref_tags && j < ref_tags->nr_tags; j++)
+                               cmpxchg(&ref_tags->rqs[j], rq, 0);
                        tags->static_rqs[i] = NULL;
                }
        }

Hi Bart,

What prevents blk_mq_tagset_busy_iter() from reading hctx->tags[...]
before the cmpxcg() call and dereferencing it after blk_mq_free_rqs()
has called __free_pages()?


So there is nothing in this patch to stop that. But it's pretty unlikely, as the window is very narrow generally between reading hctx->tags[...] and actually dereferencing it. However, something like that should be made safe in patch 2/3.

Thanks,
John


Reply via email to